IEEE TRANSACTIONS O N COMMUNICATIONS, VOL.

COM-20,

NO.

2,

APRIL

1972

115

Optimum Frame Synchronization

Abstract-This paper considers the optimum method for locating a sync word periodically imbedded in binary data and received over the additive white Gaussian noisechannel. It is shown thatthe optimum rule is to select the location that maximizes the sum of the correlation and a correction term.Simulations are reported that show approximately a 3-dB improvement at interesting signal-tonoiseratios compared to apurecorrelationrule.Extensions are given to the phase-shift keyed (PSK) sync case where the detector output has a binary ambiguity and to the caseof Gaussian data.

thereceivercanmaketentativebit decisions. Section I11 gives the necessary modification for the phaseshift keyed (PSK) sync case where the bit valuesare ambiguous until after frame synchronization is obtained.Section I V containstheresults of simulations comparingtheperformance of theoptimumruleand the correlation rule. Section V gives a derivation of the optimum sync word locating rule when the data, rather than being random binary digits, are Gaussian random variables as might be the case in some pulse-amplitude I. INTRODUCTION modulationschemes. HE MOST widely used method for providing It should be emphasized t h a t our analysis applies frame synchronization in a binary signaling scheme only to the case of asyncwordperiodicallyimbedded is to insert a fixed binary pattern or sync word intoadatastream, which is theusualcaseinspace periodically into thedatastream. On theassumption telemetry.Specifically,itdoesnotapply to the onethat symbol synchronization has already been obtained, shot synchronization problem where the sync word the receiverobtainsframesynchronizationbylocating is prefixed tothedatastreamand is itself preceded the position of the sync word in the received data either by no signal or by a periodic 1-0 pattern. It restream. mains as an interesting open problem to find the optiI n his pioneering work [ l ] on frame synchronization, mumsynchronizationruleforthisone-shotcase. Barker assumed thatthesync wordwouldbe located by passing the received digits through a pattern recogIZ. DERIVATION OF THE OPTIMALSYNC-WORD nizer, which was simply a device t o correlate successive LOCATING RULE L-digitsegments of the receivedsequence withthe LLet N denote the frame length, i.e., each L-digit sync digit sync word. The segment giving the nlaximum wordisfollowed by N-L random binary-data bits. We correlation wouldbe taken as the location of thesync assume that the receiveris to process an N digitspan word. Virtually all subsequent work on frame synchroniof the received sequence in order to locate the sync word zation has assumed this same correlation decision rule, containedtherein. I f n suchspansareactuallytobe perhaps for simplicity and perhaps in the belief that this used, the problem reduces to the above for a frame decision rule was optimal. I n hisencyclopediccoverage length of nN digits and a sync word of length nL. of synchronization, Stiffer 12, pp. 499-5021 recognizes Let T = (yo, . . . , r N - l ) denotethereceivedspanto thatthedata surrounding the sync word should be be processed where each rr is the detector output over taken into account by an optimal decision rule, but inone of the assumed-known bit intervals. The sync wordis dicates that the analysis becomes intractable and that a priori equally likely t o begin in any of the N positions theresultingtrueoptimal decision rule would be imof T . We will consider digit ro t o follow digit r N - 1 so as t o practical to implement. account for the case when the sync word begins somewhere In this paper, we derive the optimal decision rule in the last L-1 digits of T and ali subscripts on received for locating the sync word on the additive white Gausdigits will hereafterbetakenmodulo N . For example, sian noise channel and show that the effect of the data r N + Z is the digit rz. ismerelytoaddacorrectiontermtothecorrelator Let s = (so, sl, . . . , s[,-,), where each si is either +1 output so that the optimum rule is nearly as simple to or -1, be the sync word and let d = ( d L , a,,,, . . * , implement as the ordinary correlation rule. This derivadenote N-L random data bits where the d i are statistically tionis given inSection I1 for the standard case where independent random variables satisfying P r [di = +1] = Pr [di = - 1 1 = 3. Consider next the concatenation Paper approved by the Communication Theory Committee of sd = (so, s l , . . . , sL-,, d L , . . . , Let T bethe cyclic the IEEE CommunicationsSociety for publicat,ion without oral shiftoperator defined by T(sd) = ( d N - 1 , so, . . 1 S L - 1 , presentation. This work was supported by the National Aeronautics and Space Administration under NASA GRANT NGL 15- d,, . . . , dN-,J. I f the sync word actually begins in digit 004-026 a t the University of Notre Dame in liaison with the Godrm of T , we can express the received segment as .dard Space FlightCenter.Manuscript received August 16, 1971.
The author is with the Department of ElectricalEngineering. University of NotreDame,NotreDame,Ind. 46556. He is on leave a t the Laboratory for Communication Theory, Royal Technical University of Denmark, Lyngby, Denmark.
T =

4 3 T(sd) + n

(1)

where each received digit would have value either

+ z/E

116

IEEE TRANSACTIONS ON COMMUNICATIONS, APRIL

1972

or - fi in the absence of noise, and where n = (no,nl, . . . , nN-J is thecontribution of theadditive white Gaussian noise to the detector outputs. The components ni are statistically independent Gaussian random variableswith 0 meanandvariance N 0 / 2 where N o is the one-sidednoise spectral density. Let e = ( p o , p l , . . , p N - J denotetheactualvalueassumedbytherandom vector r. Then the optimum [in the sense,-of maximizing the probability of correctly locating the sync word] decision rule is to choose the estimate of m as the,value p, 0 5 p < N , whichmaximizes X, = Pr [m = p I r = e], which by the mixed Bayed rule [3, p. 751 for events and random variables becomes SI = pr(p m = p ) P r [m = p ] / p r ( p ) . Hereandhereafter, lowercase p denotesthe density function of the subscripted random variable. Since Pr [m = p ] = 1/N for all p , we may equivalently maximize S , = p r ( e I m = p ) . Letting 6 = ( 8 L , 6 , + , , . . . , where each 6; is either 1 or - 1, denote a possible value of the random data vectord , we have

is a sum over all components of e and hence is independent of p, we may subtract this sum from S , without affecting the maximization t o give
L-1
L-I

X, =
i=O

2 d z p i + , ~ / N oi=O

log, cash ( 2 d E ~ i + , , / N o )

as the quantity to be maximizedbychoice of p in the optimum decision rule. Slightly rewriting, we summarize as follows. OptimuanRuleforLocatingthe Sync W o d : Given the received segment p, taketheestimate of thesync word location m to be the value of p , 0 5 p < N , which maximizes the statistic

where

S,

=
all

pr(p
=

Id

6, m =

p)

Pr (d

= 6).

(2)

f(x)

( N 0 / 2 d E )log, cosh ( 2 d Z s / N o ) .

(6)

Since Pr (d maximize

6)

2-N-Lfor all 6, we may equivalently

S,

=
all

PI,

which uponmakinguse

of (1) becomes

By the Gaussian assumption on n, we have

Substitutingthisexpressioninto (3) andremovingall factors independent of p, we may equivalently maximize

(4) noting that each Carrying out the summation in takesononlythevalues +1 and -1, we obtain

ai

It should be noted that the first summation in (5) is theordinarycorrelation.Thesecondsummationrepresentsakind of voltage or energy correction required to accountfortherandomdatasurroundingthesync word. It should also be clear thattheoptimumstatistic S is nearly as easy to calculate as the correlation above,particularlyinthepracticalcasewherethedetectoroutputsarequantizedto 8 or 16 values (3- or 4-bitquant,ization.) Inthiscase,onlyasmallnumber of values of the function f in (6) need be stored for use in forming the correction term. Additional insight into the nature of the optimum statistic S canbegainedbyexaminingitsforminthe limiting cases of very high and very low signal-to-noise ratios. When E / N o > > 1, the argument of the coshin (6) is muchgreatert.han 1 withhighprobability so that we mayapproximate cosh (y)as (4) elf[. Using this approximationin (6), we obtain
L-I

x,

[E
,=O

e 2 f i ~ i + p s ~ / N o

] E 2 cosh (2 v% p i + , / N f l ) .
i=L

s=

c
i=O

L-1

&Pi+,, -

i=O ( P i + , , l .

Taking logarithms,

we can equivalently maximize

i=O

Noting that

Note that whenever si and pi+s agree in sign, their contributiontothefirstsummationin (7) isexactlycancelled bytheterm -Ipi+el in the second summation. Thus, only negatively correlated terms contribute to the statistic S andtheoptimum decision rule reduces to choosing that location p for the syncword that yields the least total negative correlation. When B / N o < < 1, the argument of the coshin (6) is muchsmallerthan 1 with high probability so that we may closely approximate log, cosh (y) ley the first term in its Maclaurin series expansion, namely (t) y2. Using

MASSEY: OPTIMUM

FRAME SYNCHRONIZATION

117

this approximation in (6), we obtain

=
i=O

sjpi+, -

4E No

L-l

c
i=o

PLCP2

as the statistic to be maximized by the optimum decision rule. From the form. of the second summation in ( 8 ) , we see that the correction term is anenergycorrectionin this small signal-to-noise ratio case.

ficient empirical data for a performance comparison for channel signal-to-noise ratios E / N o near unity which is the range of practical interest in space telemetry. T o conform to usual communications practice, the detector outputs ri in the simulation were quantized to 16 levels, and it was verified that 8 levels gave essentially the same results. The quantized valuesused were
(2j - 1) 1/E/6,
-8

<j 5

8,

(12)

111. PSK FRAME SYNCHRONIZATION

As Stiffler has noted [2, p. 3721, when a binary PSI< signalisdemodulatedusing a carrier reference derived from the modulated signal, there is a binary ambiguity in the detector output. With probability +,the detector output will be r i = ti ni where ti is the transmitted signal (4% si or d i ) , andwithprobability 3 the detectoroutput will be ri = -ti - ni.Whenthisambiguity is included in the analysis, a derivation very similar to that in Section 11, thedetails of which will be omitted here, leads to the following. O p t i m u m Rule for PSK Frame Sync: Given the receivedsegment e, take the estimate of the sync-word location m to be the value of p , 0 5 p < N , which maximizes the statistic

4z

L-1

log, cosh (P) i=O

log, cosh

( 2 4 pi+,,/NO) (9)

where
L-1

P = (2dz/NO)
i=O

sipi,,.

(10)

A simpleapproximateformforthisstatisticisreadily obtained. The correlation sum P can be expected to be quitelarge in general so thatthe approximation log, cosh ( P ) = IPJ - log, 2 will be quite accurate. Using this approximation, we obtain as the statistict o be maximized

X =

L-1

sipi+, i=O

L-1

c
i=O

f(Pi+,)

where f ( ) is given in ( 6 ) . We note that the usual correlationrulefor PSK framesynchronization is just to choose p to maximize the first summation in (11). Again we see the optimal decision rule adds a correction term to this usual statistic. IV.SIMULATION RESULTS The basic question remaining is whether the optimum decision rule for frame synchronization provides significantlybetterperformancethantheordinarily used but suboptimum correlation method. A theoretical comparison of performances is ruled out by the complicated relationship between probability of incorrect synchronization and channel signal-to-noise ratio. For this reason, a Monte Carlo simulation was performed to obtain suf-

and the quantization boundaries were taken halfway between adjacent quantization values. Simulations were performed using several different sync words and frame lengths as well as signal-to-noise ratios. Table I shows the results of standard frame synchronization by both the optimumandthecorrelationrulesfortwoBarkersequences and one Neuman-Hofman [4] sequence as sync words and for three different signal-to-noise ratios. The same noise and data sequences wereusedwith all sync sequences.Inspection of thistablerevealsa3-dBadvantagefortheoptimumdecisionrulefor E/No about 1 asseenbythefactthattheoptimum decision rule performs the same a t E / N o = 1 as does the correlation rule a t E / N o = 2. (This quite precise gain of 3 dB was obtained in all the simulations performed that cover a wide range of sync-word lengths and frame lengths.) The gain becomes less as the channel becomes more noisy. It should be notedfromTable I that the NeumanHofmansequenceoutperformstheBarkersequence of the same length as a sync word. The Neuman-Hofman sequencewasdesigned so asto maximizeperformance with a correlation decision rule. This simulation also shows that it is a good choice for a sync word to be used with the optimum decision rule. Table I1 shows the results of PSK frame synchronization for the same sequences and signal-to-noise ratios as in Table I. The general nature of the results are the same with the optimum decision rule a t E / N o = 1 showing the same 3-dB improvement over the correlation rule. It should be mentioned thattheoptimumrule usedin thesimulationsreportedinTable I1 wasactuallythe rule using the approximately optimum statistic S of (12). This decision ruleisindiscerniblefromthetrue optimum decision rule.

V. GAUSSIAN DATA
Forthesake of completeness, we nowconsider the case when the data digits di, rather than being limited to values of f l and -1, are instead statistically independent zero-mean Gaussian random variables with variance unity.Suchasituationmightdescribedigitized voice transmissionwithpulse-amplitudemodulation (PAM). To derive the optimum sync-word location rule, we note that the analysis inSection I1 up to (2) isunchanged

118

COMMUNICATIONS, IEEE TRANSACTIONS ON

APRIL

1972

PERCENT.4GE OF

TABLE I ERRONEOUSLY SYNCHRONIZED FRAMES I N 100 SYNCHRONIZATION TRIALS USING (COR) RULES FOR LOCATING THE SYNC WORD E/No = f OPT COR
0.31
-1, 1)

T H E O P T I M U M (OPT) AND THE CORRELATION

Sync Word

E/No OPT
0.09

= 1

COR

E/No OPT
0 .oo 0 .oo
0.09

= 2

COR

L = 13. N = 91 Barker seauence s =(l,l,l,l,l,-1,-l,l,l,-l,l,-l,l,) L = 13, N = 91 Neuman-Hofman sequence s = (-1, -1, -1, -1, -1, -1,1,1, -1, 1, -1, L = 7, N = 28 Barker sequence -1, s = (1, 0.321, 0.21 1, 0.45 -1, 0.40 -1, -1,)

0.42 0.32 0.07 0.18

0.19

0.08
0.07

0.28

0.22

P E R C E N T A G E OF

TABLE I1 ERRONEOUSLY SYNCHRONIZED FRAMES I N 100 TRIALS FOR PSK FRAME sYNCHRONIZ.4TION USING THE OPT -4ND T H E COR RULES FOR LOCATING THE SYNC WORD
E/No OPT
0.47
-1, 1) 0.49
=

Sync Word

COR
0.27 0.24 0.46

E/NQ OPT
0.14 0.14 0.37

COR

E/NQ = 2 OPT COR

L s L s L s
/

= 13, N = 91 Barker sequence =(l,l,l,l,l,-l,-l,l,l,-l,l,-l,l) = 13, N = 91 Neuman-Hofman sequence = ,(-1, -1, -1, -1, -1, -1, 1,1, -1,-1,1, = 7, N = 28 Barker sequence = 1, -1, 1, 1, -1, -1, -1)

0.39 0.39 0.63

0.12 0.13

0.63

0.21

0.40

and (2) becomes replaced by

the statistic

Using the assumption that the components of n are statistically independent Gaussian random variables with mean 0 and variance N 0 / 2 and that the components of d are statistically independent Gaussian random variables with mean 0 and variance 1, we obtain

Once again we see that the statistic to be maximized by the optimum sync-word location rule consists of the ordinary correlation together with a correction term. For Gaussian data, the correction term is again seen to be a true energy correction. VI. SUMMARY This paper has considered the problem of obtaining an optimal estimate of the location of the sync wordin a dataframeforbinarydatainthestandardcaseand in the PSK sync case and alsofor Gaussian data samples. I n all cases it was shown that the optimum decision rule is to maximize a statistic that is the sum of two terms, the first being the usual correlation and the second being anenergycorrectionthattakesintoaccountthefact that the sync word is imbedded in data. It was verified by simulation that the optimum rule provides about a 3-dBadvantageovertheordinarycorrelationrulein the interesting case of signal-to-noise ratios E / N o near unity. It was also noted that the optimum decision rule is only slightly more complicated to implement than the correlation rule. VII.ACKNOWLEDGMENT The author is grateful to C.Seguin for programming thesimulationsreportedinSectionIVandtothereviewers of this paper for several helpful comments.

Carryingouttheindicatedintegrationsanddropping factors independent of p, we may equivalently maximize

i=O
L-1

c
i=O

&Pi+

-N O

+ 2E c
N-l
%=L

Adding the quantity

which is independent of p, we may equivalently maximize

IEEE TRANSACTIONS ON COMMUNICATIONS, NO. VOL. COM-20,

2, 119 APRIL 1972

REFERENCES
111 R. H. Barker, Group synchronization of binary digital systems, in ~Commwnicatiom Theory, W. Jackson, Ed. London. Butterworth. 1953. an. 273-287. [21 J. J . Stiffler, TheorydjSynchronous Comnmnications. EnglewoodCliffs, N. J.: Prentice-Hall, 1971. [31 J . M. Wozencraft and I. M. Jacobs, Principles of Communication Engineering. New York: Wiley, i965. [41 F. Neuman and L. Hofman, New pulse sequences with desirable correlation properties, in Proc. Nat. Telemetry Conj., Washington, D.C., Apr. 1971, pp. 27-282.
James L. Massey (S54-M55-F71) was bornin Wauseon, Ohio, on February 11,1934. He received the B.S.E.E. degree from the University of Notre Dame, Notre Dame, Ind., in 1956, and the

S.M. and Ph.D. degrees in electrical engineering from the Massachusetts Institute of Technology, Cambridge, in 1960 and 1962, respectively. From 1956 t o 1959 heserved as a Communications Officer in the U. S. Marine Corps. Since 1962 he has been on the Faculty of the University of Notre Dame where he is now Professor of ElectricalEngineering. I n 1966-1967, he was a Visiting Associate Professor of Electrical Engineering at M.I.T. He is spending 1971-1972 as a Visiting Professor at the Royal Technical University of Denmark, Lyngby. Sigma Xi, Tau Beta Pi, Dr. Massey is a member of Eta Kappa Nu, and the American Society for Engineering Education. He is a past chairman of the IEEE Group on Information Theory.

Performance o f a First-Order Transition Sampling Digital Phase-Locked Loop Using Random-Walk Models

Abstract-A new mechanization of a first-order all digital phaselocked loop (ADPLL) is discussed and analyzed. The purpose of the loop is toprovide continuous tracking of theincoming waveform corrupted by the presence of white Gaussian noise (WGN). Based on a random-walk model, solutions are obtained for the steady-state timing-error variance and mean time to slip a cycle.A s a resultof the mean first-time-to-slip analysis, difference a equation and its solution are obtained that generalize a result of Feller [l]. Using a procedurethatappears new, an upperboundon the timing-error bias due to a Doppler shift of the synchronized waveform is also derived. An example, for which the results presented here are applicable, is considered in some detail.

I. INTRODUCTION
NALOG phase-locked loops (PLL) have long played an important role in modern communication systems. Their theory of operation is well documented in numerous papers throughout the past 15 years.The increased inhest in digital communication systems,dueprimarilytothedecreased size and cost and increased reliability, led to the digital phase-locked loop ( D P L L ) . One of theearliestreportedDPL,Lwas
Paper approved for publication by the Communication Theory Committee of t h eI E E E Communications Society after presentation at the 4th Hawaii International conference on System Science, January 1971. Thispaper presents the results of one phase of research carried out at the Jet Propulsion Laboratory, California Institute of Technology, underContract NAS 7-100, sponsored by the National Aeronautics and Space Administration. Manuscript received July 1, 1971. Theauthor is with the Jet Propulsion Laboratory,Pasadena, Calif. 91103.

constructed by adding a samplc and hold circuit between the filter and the voltage-controlled oscillator (VCO) [2] inananalog loop. Withtheadvent of integrated circuits,more refined and morenearlyall-digital PLL emerged [3], [4]. These loops utilize analog integration for the midphase in-phase (MI) type phase detector, with the remaining functions of the loop beingaccomplisheddigitally. A loop reported by Gota [5] was, except, for the use of a digital t o analog converter, all digital in opemtion; he did notconsider noise in the analysis of hisloop. To date the most common function of theDPLL has been to providesynchronization of asignal.However, this is not the only application, for example, an alldigitalphase-locked loop (ADPLL)has been builtfor FM demodulation [6]. It was designed for potential application in large multiple data set. installations, which provide low-speed serial data communications, for timeshared computers. Additional applications of D P L L are summarized in a paper by Gupta 171. RecentlyanADPLL,employinga newsimple type of phase detector has been reported [8] for square-wave signal waveforms. The loop was conceived toprovide tracking of the subcarrier signal for a command system. Basically, tracking was accomplished by 1) sampling the input waveform a t t h e points in time where the signal transitions or axis crossing occur, 2) accumulating m of these samples, and 3) incrementing the phase of the local reference (clock) in such a direction as to bring the value